The Asset Outperformer Engine's stated construction is a decile rule: the top decile becomes the long book and the bottom decile becomes the short book. That is a clean specification and it hides an expensive problem. A hard cut applied to a continuous score means the membership of your book is decided by whichever names happen to sit within a rounding error of the boundary on scan day, and those names change constantly without the underlying picture changing at all.
Where the churn actually comes from
Two properties of the ranking combine badly. The first is scan frequency: the engine runs a full scan every six hours across the universe. The second, and the more important one, is that the score distribution is dense near the top rather than well separated.
The captured board makes this concrete. On the crypto tab, five names sat at exactly 69 and a further five at exactly 68. On the US stocks tab, six separate names carried a score of exactly 70. These are integer scores on a 0 to 100 scale, so ties are not rare edge cases, they are the normal condition in the region where the cut line falls. A boundary drawn through a block of names sharing one integer will reshuffle on the next scan for reasons that have nothing to do with signal.
Compounding this, the score is a composite of seven timeframes, and the shortest of those are 1, 3 and 7 day windows. Those components turn over fast. An asset whose 30, 90, 180 and 365 day standing is completely unchanged can move several points on short-horizon noise alone, which is enough to cross a boundary that a dozen names are already sitting on.
Measuring the churn before trying to fix it
Damping rules are cheap to propose and expensive to adopt blind, so instrument first. Three measurements are enough to size the problem.
The first is boundary residency. For each rebalance date, record every name within a stated score band either side of the cut, and track how many of them change side at the next observation without moving more than a point or two. This isolates churn caused by the boundary from churn caused by real rank movement, and it is the number that justifies the whole exercise.
The second is name survival. Take the decile membership on a formation date as a cohort and record what fraction remain in the decile after one week, two weeks, a month. A cohort that empties fast tells you the rebalance interval is shorter than the signal's persistence, which is a different problem from boundary noise and needs a different fix.
The third is realised turnover, expressed as the fraction of the book replaced per rebalance and annualised. Note that this must be measured on the population you actually trade, not the combined board. The module separates crypto, stocks, FX, precious metals, commodities, ETF and index, and mutual funds into type tabs, and churn behaves differently in each because the underlying volatilities differ by an order of magnitude.
What each damping rule costs
Three families of fix are available and they trade off differently. None is free, and the cost of each shows up somewhere other than turnover, which is why they are worth comparing rather than picking on instinct.
A buffer zone holds an existing position until its rank falls past a wider exit boundary, so a name enters at the top decile and leaves only at, say, the top quintile. This is the cheapest to implement and the easiest to explain in a review. It cuts boundary churn sharply because a one-point wobble no longer triggers anything. Its cost is a book that is systematically staler than the ranking, holding names the engine no longer considers top decile, and in a fast-decaying signal that stale tail is where the drag accumulates.
Rank hysteresis makes entry stricter than exit rather than looser, requiring a name to clear the boundary by a margin before it is added. It reduces the number of positions taken on noise, and it costs you the fastest movers, because a name accelerating through the boundary is exactly the case a margin requirement delays. If the strategy earns its return from early entry, this is the wrong lever.
A minimum holding period simply refuses to close a position before a stated number of days. It gives the most direct control over turnover because it caps it arithmetically, and it is the most brittle in a drawdown, since it holds through deterioration by construction. Anyone adopting it needs an explicit risk override that sits outside the ranking, and needs to write down in advance what triggers it.
The comparison to run is turnover reduction against forgone return, measured on your own history rather than assumed. Every one of these rules trades some of the ranking's responsiveness for stability, and the right amount depends on where your costs actually bind.
Fitting the damping rule to the mandate
The choice is not purely empirical, because the constraint that binds hardest is usually capacity rather than accuracy.
A book small enough to trade the full decile at will can afford a responsive rule and should probably use a modest buffer and nothing else. A book large enough that entering a position takes multiple sessions cannot use a rule that reverses within a week, and for that book a minimum holding period is not a damping preference but a structural requirement. The market cap column is the place to check this: at capture, ranked names ran from 1.79 M USD up to 2.38 B USD on the crypto tab alone, so a single decile can contain positions that differ in tradable size by three orders of magnitude.
Cadence is the other half of the fit. The engine rescans every six hours. Adopting that as a rebalance interval imports the maximum possible amount of boundary noise, and for most institutional books the correct response is to sample the ranking at a frequency the book can actually act on and treat intermediate scans as information rather than instruction.
What to log so the rule survives review
A damping rule changes which positions you hold, which means it will eventually be blamed for one. The defence is a record made at decision time rather than reconstructed afterwards.
Log the rule parameters and the date they were last changed, because a buffer width that has been quietly retuned three times is an overfitting problem wearing a governance problem's clothes. Log the score and rank of every position at entry and at every rebalance, including the ones the rule caused you to keep, so the counterfactual book is reconstructible. Log the scan timestamp each decision was taken from. And record, separately, the names the rule excluded and how they subsequently performed, because that is the forgone return term and it is the only honest way to know whether the damping is paying for itself or simply making the turnover report look better.